Helpful AI Models - You Can't Always Get What You Want, But You Might Get What You Need

Attend this plenary conference talk exploring how AI models should be optimized for human-computer collaboration rather than just accuracy or user satisfaction. Learn about three research examples demonstrating effective human-AI teamwork: vocabulary learning through adaptive flashcard scheduling that combines perceived and actual helpfulness, strategic negotiation assistance in the board game Diplomacy using grounded statement analysis and value functions, and collaborative fact-checking where computers help humans identify false claims while avoiding overconfident errors. Discover how measuring and optimizing human-computer workflows can lead to more effective AI systems, with insights into balancing human versus computer skills and appropriate evaluation datasets. The presentation, delivered by Jordan Boyd-Graber from the University of Maryland, draws from extensive research in human-centered AI applications including topic modeling, question answering, and machine translation, offering perspectives on creating AI systems that truly enhance human capabilities rather than simply replacing them.